The lessons that AI should take from the 1950’s Highway Act—and the disastrous pitfalls to avoid

In 1956, the U.S. Congress introduced the Federal Aid Highway Act, the largest public works project in American history to that point. It called for the construction of 41,000 miles of the Interstate Highway System, with the federal government paying 90% of the costs. “Together, the united forces of our communication and transportation systems are dynamic elements in the very name we bear—United States,” said President Dwight Eisenhower. “Without them, we would be a mere alliance of many separate parts.”

Despite the lofty intentions, federal highways had adverse consequences, bulldozing entire urban neighborhoods and encouraging suburban sprawl while turning the U.S. into a traffic-filled, car-dependent country. Today, the United States continues to spend billions to fund and expand highways, even as other countries pursue ways to cut back on car use through public transportation and more active ways of getting around, like walking and cycling.

A 70-year-old infrastructure project may seem like an unlikely corollary to our modern tech landscape, but in many ways, the Federal Aid Highway Act provides a valuable lesson for one of the biggest issues facing the tech industry today: How will we produce enough power to support the growing number of AI-focused data centers? To echo the words of President Eisenhower, in bringing together the dynamic communication and transportation systems around AI, will we lay the foundation for a united future, or will we find ourselves a mere alliance of many separate parts?

AI’s immense potential has united the public and private sectors around a concept called Sovereign AI. Defined as a nation’s capabilities to produce AI using its own infrastructure, data, workforce, and business networks, Sovereign AI has raised its profile enough to be highlighted in a recent earnings call by AI powerhouse NVIDIA. Japan, France, Italy, and Singapore were among the companies mentioned as investing hundreds of millions of dollars in their AI ecosystems.

Along with building the technology, nations are also grappling with the growing need to feed AI’s voracious power demands. According to the World Economic Forum, the computational power needed to sustain AI’s rise is doubling roughly every 100 days. Additionally, the energy required to run AI tasks is already accelerating, with an annual growth rate between 26% and 36%. This means by 2028, AI could be using more power than the entire country of Iceland used in 2021.

This is where the highway model comes into play. We could choose to find ways to supply more energy to power AI (build more highways), or we could discover how to lower AI energy costs (invest in high-speed rail). One path leads to a power-sucking, climate-destroying future, while the other is sustainable and profitable.

The good news is that there’s already movement on several fronts that demonstrates dramatically lowering energy costs at the source—the AI data center itself—is possible. This approach doesn’t just make AI more affordable; it aims to fundamentally reduce energy consumption per AI operation. By improving efficiency at the hardware and software levels, we can process more AI tasks with less energy, rather than simply enabling more usage at a lower cost. This efficiency-first approach will help large-scale national AI efforts and opens its potential to government agencies, lower-margin industries, and smaller companies by making it both more affordable and more sustainable. The goal is to break the cycle of increased usage leading to increased energy consumption, instead focusing on doing more with less.

Researchers like Sara Hooker are advocating for a centralized rating system that evaluates the energy efficiency of AI models, similar to how cars are rated for energy standards. At the same time, the MIT Lincoln Laboratory Supercomputing Center is developing techniques like setting limits on how much power certain components can draw and tool that stop AI training earlier than usual, all with an eye toward finding ways to reduce power, train models efficiently, and make energy use transparent. We can also look to Europe, where more than 100 data center operators have committed to making their locations climate-neutral by 2030. It’s still early, but last year CyrusOne became the first company to have all its data centers comply with the Climate Neutral Data Centre Pact’s reporting terms.

One approach that I’m particularly passionate about is finding a way to eliminate CPUs (central processing units) altogether in AI Inferencing servers. (In the AI world, training is the stage when the model is still learning how to draw conclusions. Inference is when the model is put into action.) Eliminating this major bottleneck in running trained AI models would significantly improve AI data center efficiency and performance

Otherwise, it’d be like building all these faster sports cars and AI-enabled SUVs but using the same old roads without sensors, signals, or satellite data to tell those vehicles the best, most efficient routes on any given day or hour—or that there’s an upcoming vehicle in the wrong lane.

It’s remarkable that the same underlying CPU infrastructure that powered our PCs and the Internet Era now hinders progress in the AI Age. My company is developing one potential solution—a new standard design that circumvents the CPU to make AI chips more efficient. Others are building chips specifically designed to process AI that have already been deployed and are touted as faster and cheaper than those currently used.

Efforts like these demonstrate the possibility of a sustainable, affordable AI future, with the potential to solve the extraordinary cost and complexity problems of running AI models and proven benefits in energy efficiency and better performance. For the future of AI, we can either invest heavily in outdated ways of supplying power that put an additional strain on our current power grids or find a way to lower costs at the source—the AI data center itself—with baked-in systems engineering that does most of that heavy work.

If we had the chance to do it all again, would we choose a car-heavy approach or the bullet train?